As a practitioner in the field and in similar circumstances, the answer will probably be: it doesn't really matter. Personally I prefer PyTorch, but I can work with whatever the client already uses, if anything. For one of my clients I worked in both because there was no viable deployment option out of PyTorch to one of their platforms. There still isn't, so both models are still maintained and trained.

If they aren't 100% set in their ways, I do make them aware that things will move at about half the speed with TF, so they'll effectively be paying twice as much. If they are set in their ways, I do not mention it, since I'm not going to change their mind anyway.

That said PyTorch 1.5.0 is just broken pretty much - tensor permute (which in computer vision you end up doing for every input tensor) is 10x slower than it used to be. There's an issue in GitHub already.

I'm beginning to worry about PyTorch.